OttoAI

A local-first, extensible coding agent for your terminal — think "Copilot CLI, but yours." Runs on local models (Ollama / Qwen3) by default, works fully offline, and can optionally use cloud providers (OpenAI, Anthropic, Gemini, Azure, OpenRouter, or any OpenAI-compatible endpoint) when you choose to turn them on.

Features

• 🧠 Local by default — Ollama + Qwen3 out of the box, no cloud required.
• 🔌 Cloud-optional & pluggable — flip to online and use OpenAI / Claude / Gemini / Azure / OpenRouter / any OpenAI-compatible server via config.
• 🛠️ Real agent tools — shell, file read/write/edit, grep/glob code search, git, web fetch, cross-session history search, and MCP (Model Context Protocol) servers.
• 💾 Sessions + memory — every conversation is persisted in SQLite, with FTS5 keyword search and optional embeddings for semantic recall across past sessions.
• ♻️ Conversation compaction — long chats are automatically summarized to stay within context, while the full history stays searchable.
• 🧩 Extensible — add providers/tools via pip entry points or drop a .py file in ~/.ottoai/plugins/.

Install

pip install ottopilot        # from PyPI (once published)
# or, from source:
pip install -e .

Install Ollama and pull the default model (light, runs on ~8GB RAM):

otto pull qwen3:1.7b

On 16GB+ machines you can use a stronger model, e.g. otto pull qwen3:8b or otto pull qwen2.5-coder:7b, then set it via /model or default_model.

Usage

otto                      # interactive REPL (local-first)
otto chat "explain this repo"
otto chat --provider openai --model gpt-4o "review my diff"
otto models               # list models for the default provider
otto catalog              # installable models + specs (size, RAM, best-for)
otto catalog --ram 8      # only models that fit 8GB RAM
otto catalog --coding     # only coding-focused models
otto sessions             # list past sessions
otto resume               # resume the latest session
otto search "auth bug"    # search across all past sessions
otto config init          # write ~/.ottoai/config.toml to customize

In the REPL

/model <name>        switch model
/catalog [ram|coding] installable models + specs (e.g. /catalog 8)
/provider <name> [m] switch provider
/online  /local      toggle cloud access
/stream              toggle streaming output
/think [on|off]      toggle model reasoning (faster vs. higher quality)
/search <query>      recall past sessions
/remember subj | fact store a durable fact
/compact             compress the current conversation
/tools  /sessions  /memory  /help  /quit

Configuration

Config lives at ~/.ottoai/config.toml (run otto config init). Highlights:

local_only = true          # set false (or /online) to allow cloud providers
default_provider = "ollama"
default_model = "qwen3:1.7b"  # light default; bump on 16GB+ machines
stream = true              # token-by-token output (/stream to toggle)
think = true               # model reasoning: true = higher quality, false = faster (/think)

[compaction]
enabled = true
trigger_messages = 40
keep_recent = 12

[embeddings]
enabled = false            # set true for semantic cross-session search
provider = "ollama"
model = "nomic-embed-text"

[providers.openai]
type = "openai_compatible"
base_url = "https://api.openai.com/v1"
api_key_env = "OPENAI_API_KEY"

[mcp_servers.filesystem]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "/path"]

Extending OttoAI

Local plugin — create ~/.ottoai/plugins/my_tool.py:

from ottoai.tools import Tool

class HelloTool(Tool):
    name = "hello"
    description = "Say hello"
    parameters = {"type": "object", "properties": {"who": {"type": "string"}}}
    def run(self, who="world", **_):
        return f"Hello, {who}!"

TOOLS = [HelloTool]
# PROVIDERS = {"my_type": MyProviderClass}  # providers work the same way

Pip plugin — expose ottoai.tools / ottoai.providers entry points in your package.

License

MIT